Unsupervised Segmentation and Quantification of COVID-19 Lesions on Computed Tomography Scans Using CycleGAN

ABSTRACT

Rationale Lesion segmentation is a critical step in medical image analysis, and methods to identify pathology without time-intensive manual labeling of data are of utmost importance during a pandemic and in resource-constrained healthcare settings. Here, we describe an unsupervised method of automatic lesion segmentation and quantification of COVID-19 lung tissue on chest Computed Tomography (CT) scans. Methods: Anonymized human COVID-19 (n=53), and non-pathologic control (n=87) inspiratory CT scans were used to train a publically available cycle-consistent generative adversarial network (CycleGAN), to convert the COVID-19 CT scans into generated "healthy" equivalents. Difference maps were created by subtracting the Hounsfield Units (HU) value for each voxel in the generated image from that of the original COVID-19 image. We then used these difference maps to construct 3D lesion segmentations to further quantitatively characterize COVID-19 lesions in an automated pipeline. Results: The CycleGAN produced lesion segmentations from COVID-19 CT scans of varying radiologic severity ranging from cases of patchy ground glass opacities to diffuse consolidative lesions. Images of COVID-19 patients showed higher HU intensity in original vs. generated images at sites of pulmonary lesions, while preserving normal parenchyma, fissures, vasculature, and airways (Figure 1, upper panels). The generated images showed larger lung gas volumes and lower tissue masses compared to their corresponding original COVID-19 images (p<0.001). Subtraction of the generated images from their corresponding original COVID-19 CT scans yielded difference maps showing the pathological tissue alone (Figure 1). Control, non-pathologic CT images were given as input to the CycleGAN, resulting in generated images nearly superimposable with the originals with no difference in gas volume or tissue mass (Figure 1, lower panel). Conclusions: To our knowledge, this is the first unsupervised COVID-19 lesion segmentation approach. Our automated lesion model performed well in mild and severe COVID-19 cases without the need for manually labelled lung segmentations as inputs. An automated lesion segmentation model can be used clinically to rapidly and objectively quantify pathologic pulmonary tissue to inform disease prognosis and treatment. Automated radiologic techniques, such as our model, circumvent the traditional bottle-neck of manually labeling data which has limited the scale and thus the impact of quantitative radiologic medical research.